cross-encoder-reranking

A production-grade RAG system and neural voice assistant for low-resource Kannada literature ('Heli Hogu Kaarana'). It features BM25 + ChromaDB hybrid search with page-level routing, Gemini/Groq model fallbacks, and Sarvam AI TTS for streaming audio. Validated using RAGAS with 0.92 Faithfulness and 0.88 Answer Relevancy.

Hybrid RAG pipeline (Dense + BM25 + Cross-Encoder Reranker) with conversational memory and systematic evaluation. 5 retrieval experiments, 92.3% Hit Rate, 5.0/5.0 LLM-as-a-Judge. Built with LangChain LCEL & ChromaDB.

First-place official system for JOKER 2026 (CLEF) Task 1 English: a three-stage humor retrieval pipeline pairing hybrid sparse-dense retrieval and cross-encoder reranking with a rationale-distilled LLM judge ensemble. 0.6347 MAP. Team VANGUARD.

Store embeddings, run filtered vector search, and rerank results with MMR or a cross-encoder. Built on FastAPI + Qdrant.

CortexRAG: "Memory-Augmented Hybrid Retrieval"

Successfully implemented a retrieval evaluation pipeline that tests how well a RAG system finds relevant call center transcript chunks, using contextual precision and recall as metrics.

Production-grade Medical RAG Assistant with Hybrid Retrieval (Dense + BM25 + RRF + CrossEncoder), Layered Guardrails, Developer Diagnostics, PDF Evidence Viewer, Langfuse Observability, Offline Evaluation, and Background Knowledge Processing using FastAPI, React, Qdrant, and Celery.

Multimodal RAG over PDFs. Three parallel pipelines (LLM summaries, raw atomic content, CLIP visual) with HyDE expansion and cross encoder reranking, answers cited inline.

Successfully developed a multi-domain RAG pipeline that retrieves and answers questions grounded in research papers on lung cancer detection, NSCLC microbiome classification, and Bhutan land cover, with a full safety evaluation suite covering toxicity, information leakage, and scope adherence using DeepEval.

An advanced RAG system utililzing text datasets chunked dynamically. Engineered a dual-stage retrieval pipeline using all-MiniLM-L6-v2 embeddings to pull 40+ FAISS candidates, refined via Cross-Encoders to feed the top 8-10 most relevant chunks to the LLM. Multiple vector-stores for context-specific RAG.